Visual detection and recognition using local features

Detection and recognition of objects in images is one of the most impor- tant problems in computer vision. In this thesis we adhere to a traditional bottom–up detection and recognition framework, where the objects are first localized with a sliding window detector before being identified. We make multiple contributions along this path. All of the contributions pertain to the central theme of local image features. We demonstrate improved object detection performance with our proposed feature extraction process, which generalizes the traditional feature extrac- tion methodology of pooling atomic appearance information (e.g., image gra- dients) around pixels in localized histograms. In addition, we propose a method to fuse two types of information sources in a locally discriminative manner by leveraging local class-dependent correlations. For the recognition task, we adopt a state–of–the–art metric learning method and modify it to handle unknown identities. Lastly, the computational improvements achieved through leveraging par- allelism are brought together by the Vision Video Library (ViVid), which we release as open source to the research community